Event Understanding of Human-Object Interaction: Object Movement Detection via Stable Changes - Intelligent Video Event Analysis and Understanding - page 203

region in the foreground region list (e.g. when a person places a book on something,
the human region and the topic region will match the region of the human with
the topic in a previous frame). In these cases, the human region is expected to be a
larger object region. So, the larger region is identified as the matched region, and the
smaller region is newly registered to the foreground region list. If the region stored
in the foreground region list is not accessed for a long time, the region is deleted.

4 Object Detection and Background Update

This section describes the technique for categorizing the extracted regions into non-
objects and objects via their motion history, and detecting object placement and
object removal from foreground regions' edge subtraction result.

4.1 Region State Detection

After the method tracks extracted regions, the method detects extracted regions'
motion via the partial frame subtraction technique. Fig. 5 (a) shows an overview
of partial frame subtraction. In the partial frame subtraction method, the extracted
region is divided into M 4 in our implementation), and motion

is detected in each subregions using the frame subtraction technique. If the ratio
of pixels detected moving in the subregion are above a fixed threshold, the whole
extracted region is classified as a moving region. Partial human motion (e.g. arm
movement) is expected to form a cluster of moving state pixels, so partial frame
subtraction enables to robust detection of partial motion (an example is shown in
Fig. 5 (b)).
Then the method updates the attribute of the extracted foreground regions by their
motion detection history. Each foreground region has stability value S

× N subregions (4 ×

,wherek
is an incremental parameter. The stability values are updated by the region's motion
detection result as follows.
1. If the region state is moving, S

( t )

( t )= S ( t − 1 )+ k

2. If the region state is stable, S k

An unstable result of motion detection may cause misclassification of the region's
attribute. To avoid detecting attributes under unstable motion detection, we set in-
cremental parameter k

( t )= S ( t − 1 ) −

= 2 if the region's motion detection result is equal to the

previous result, and k = 1 if the region's motion detection result is not equal to the

previous result.
After updating the stability value, then the method calculates the region's at-
tribute via a state machine driven by stability value S . Fig. 6 shows the state

diagram of the region's attribute. Three attributes are defined as follows.
1. Transition region
The initial attribute of a foreground region is the transition region. If the region
with transition region attribute does not transfer to the update-prohibited region
and the static region, the region transfers to the transition region.

( t )